General Control Law for Multi-Limb robots for pushing task, by adapting posture

Abstract

Multi-limb robots have the potential to be dexterous in manipulation and be able to locomote in challenging terrain. In the future they should match or supersede the movement capabilities of humans and animals, thus becoming useful to support humans in their daily activities or perform those activities individually in environments which are dangerous to humans. Humans and animals carry out such tasks by means of skills such as pushing, pulling, twisting and grasping, and for the multi-limb robots to perform such tasks, they also need to possess such skills. This thesis addresses the pushing skill by developing a general controller for multi-limb robots. The controller tries to execute the pushing task in a way humans would do. This controller is novel because it considers the possibility of using surrounding objects as potential support for high force generation. The resulting optimization-based motion is surprisingly similar to how humans push large objects. The controller involves three major modules. The first module computes a Posture Tree, which takes geometric information about the environment and the robot, and generates a set of possible postures. These include the selected contacts in the robot and their respective location on the object. The second module is the Posture Optimization, which uses a cost function based on stability, friction at contact, joint and torque range to optimize each posture. The final module consists of an Operational Space Controller (OSC), which selects the least costing posture and executes it by dividing the complete posture into several tasks. Each task is operated within the null space of its higher level task. The force is generated by modifying the end targets of the end-effectors (pushing contacts) and the robot resists the destabilizing moment resulting from the reaction force, by moving the center of gravity in the direction opposite to an approximation of the ZMP movement. Initial verification involving first two modules was carried out with a humanoid robot of 29 Degree of Freedom (DOF) and its results are included in this report. For final verification, a more generic multi-limb robot of 8 end-effectors (Octobot 32 DOF) and a typical living environment were conceptualized, developed and evaluated. For validation of the complete controller a number of simulations were carried out in V-Rep. Simulation results depict the controller’s ability to identify the suitable posture for any multi-limb robot, given some pushing task in any given environment. The results collectively demonstrate the controller’s ability to handle heavy object manipulation, autonomous contact planning, adapting posture according to friction at the contacts and adapting posture according to the pushing force requirement. Unlike the present controllers, the complete control module developed here is more generic both in terms of the system (robot) and the environment in which the system operates.